Scholarlink

<p>Network representation learning aims to learn low-dimensional representations of nodes while preserving the structure and attribute information of the network. However, most existing methods mainly focus on the explicit structure between nodes, neglecting the capture of latent information and the data distribution of learning network low-dimensional representations. This often results in poor performance when learning low-dimensional representations. To address this issue, an implicit association relationships perception network representation learning model based on adversarial training variational graph autoencoder is proposed. On one hand, the model can effectively capture the latent implicit association information between different nodes while preserving the topological structure and attribute information. On the other hand, the model can flexibly learn the potential distribution of representations for each node, enhancing its generalization ability in real-world network data. Specifically, a latent graph structure learning module is first proposed, which captures the implicit association relationship between nodes and constructs a relationship enhancement graph, enabling node representations to contain more latent semantic and structure relationship information. An implicit association relationships perception variational graph autoencoder module is then designed, which takes the relationship enhancement graph as input and simultaneously reconstructs the structure information of relationship enhancement graph and the global attribute information, embedding the structure, attribute, and implicit associations into the low-dimensional representations. Furthermore, an adversarial training module is also designed to further optimize the data distribution of the learned representations and enhance the robustness of the obtained node representations. Extensive experiments are conducted on six real-world datasets across three network analysis tasks. The results demonstrated that the learned node representations serve as effective features for network analysis tasks and achieved superior performance.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Implicit association relationships perception network representation learning based on adversarial training variational graph autoencoder

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Lili Han,
  • Hui Zhao

摘要

Network representation learning aims to learn low-dimensional representations of nodes while preserving the structure and attribute information of the network. However, most existing methods mainly focus on the explicit structure between nodes, neglecting the capture of latent information and the data distribution of learning network low-dimensional representations. This often results in poor performance when learning low-dimensional representations. To address this issue, an implicit association relationships perception network representation learning model based on adversarial training variational graph autoencoder is proposed. On one hand, the model can effectively capture the latent implicit association information between different nodes while preserving the topological structure and attribute information. On the other hand, the model can flexibly learn the potential distribution of representations for each node, enhancing its generalization ability in real-world network data. Specifically, a latent graph structure learning module is first proposed, which captures the implicit association relationship between nodes and constructs a relationship enhancement graph, enabling node representations to contain more latent semantic and structure relationship information. An implicit association relationships perception variational graph autoencoder module is then designed, which takes the relationship enhancement graph as input and simultaneously reconstructs the structure information of relationship enhancement graph and the global attribute information, embedding the structure, attribute, and implicit associations into the low-dimensional representations. Furthermore, an adversarial training module is also designed to further optimize the data distribution of the learned representations and enhance the robustness of the obtained node representations. Extensive experiments are conducted on six real-world datasets across three network analysis tasks. The results demonstrated that the learned node representations serve as effective features for network analysis tasks and achieved superior performance.